Computer System Including a Processor Configured to Evaluate Browse History Data and Social Media Data

The present application claims priority from and is a continuation of, U.S. patent application Ser. No. 17/334,925, filed on Mar. 31, 2021, which claims priority from and is a continuation of U.S. patent application Ser. No. 15/941,778, filed on Mar. 30, 2018, which claims priority from and is a continuation of U.S. patent application Ser. No. 14/815,796, filed on Jul. 31, 2015, now issued as U.S. Pat. No. 9,959,553, which claims priority from and is a continuation of U.S. patent application Ser. No. 14/257,394, filed on Apr. 21, 2014, now issued as U.S. Pat. No. 9,129,305 which claims priority from and is a continuation of U.S. patent application Ser. No. 13/860,461, filed on Apr. 10, 2013, now abandoned, which claims priority from and is a continuation of U.S. patent application Ser. No. 13/284,799, filed on Oct. 28, 2011, now issued as U.S. Pat. No. 8,443,384, which claims priority from and is a continuation of U.S. patent application Ser. No. 12/098,385, filed on Apr. 4, 2008, now abandoned, which claims priority from U.S. Provisional Patent Application No. 60/910,581, filed on Apr. 6, 2007, U.S. Provisional Patent Application No. 60/910,606, filed on Apr. 6, 2007, and U.S. Provisional Patent Application No. 60/910,583, filed on Apr. 6, 2007. The contents of each of these applications are incorporated by reference herein in their entirety.

Advertisements are increasingly being used by web site operators as significant sources of revenue. Advertisements are often provided in the form of banners, textual advertisements, pop-ups, or the like. When a user clicks on an advertisement, the user may be taken to a web site or other network resource that features a product or service highlighted in the advertisement.

Web site owners often contract with advertising partners to display advertisements on the web site owners' web pages. A web page on a web site might, for instance, include advertisements provided by an advertising partner's servers. The advertisements selected by the advertising partner may be randomly selected. Alternatively, the advertisements may be more targeted toward users. For example, advertisements may be selected that closely match one or more keywords listed on the web page the user is viewing or on a web page the user has viewed. Oftentimes, however, the subject matter of these advertisements does not align with the current preferences or interests of the viewer.

In certain embodiments, a system for generating targeted advertisement recommendations includes a data aggregation module that can obtain word data from social-network data of one or more network resources. The word data can include a plurality of words in the social-network data. The system can further include a relationship mining module in communication with the data aggregation module. The recommender module can create word relationships between selected ones of the plurality of words in the word data to produce relationship data, where each of the word relationships can reflect a degree of association between two or more of the selected words. This degree of association may be based at least in part on one or more of: a frequency of the selected words in the social network data, a recency of a subset of the social-network data that includes the two or more selected words, and an authority factor of the subset of the social-network data. Moreover, the system may include a recommender module in communication with the relationship module, which may access browse information of a target user to identify one or more first words in the target user's browse information, identify the one or more first words in the relationship data, identify one or more second words in the relationship data that have one or more of said word relationships with the one or more first words, and identify. one or more advertisements having at least one keyword that corresponds to the one or more second words.

Various embodiments of a computer-implemented method of generating targeted advertisement recommendations include obtaining word data from one or more network resources, where the word data includes words in content of the one or more network resources. The method may further include creating word relationships between selected words in the word data to produce relationship data, each word relationship indicating a degree of association between two or more of the selected words. In addition, in certain embodiments the method includes analyzing browse information of a target user to identify one or more first words in the target user's browse information, identifying, based on the word relationships, second words in the relationship data that are related to the first words in the target user's browse information, identifying one or more advertisements having at least one keyword that corresponds to the second words, and selecting at least a portion of the one or more advertisements from a data repository to provide to the target user.

Additionally, in certain embodiments, a computer-implemented method of determining meanings of ambiguous words includes obtaining word data from one or more network resources, where the word data includes words in content of the one or more network resources. For a subset of the word data, the method further includes determining parts of speech of selected words in the subset of the word data. In Page 3 of 36 addition, the method may include determining definitions of the selected words to determine forms of the selected words and determining one or more categories or related words for each of the word forms. For two selected word forms, the method may also include connecting the two selected word forms in one or more graphs in response to determining that categories or related words of the two selected word forms overlap, where connecting the two selected word forms in the one or more graphs may include creating weighted word relationships between the two selected word forms. Moreover, the method may include determining that a selected graph having a highest total weight of word relationships compared to one or more other graphs indicates definitions of one or more of the selected words in the subset of the word data.

Neither this summary nor the following detailed description purports to define the inventions disclosed herein. Certain inventions disclosed herein are defined by the claims.

Many drawbacks exist with advertising systems that provide advertisements that closely match keywords on a web page. One drawback is that the advertisements provided may merely duplicate content found on the web page. Such advertisements may not be interesting to a user. For example, an advertisement for a Ford® truck displayed on a web page featuring a story about trucks may not entice a user to view the advertisement because the user is already viewing content about trucks.

Another problem with keyword-based methods occurs when advertising partners have small sets of advertisements. In such instances, the keywords on a given web page might correspond poorly to the advertisements in the set because the set is too small to have an advertisement for every possible keyword. The advertisement provided to the user may therefore be so disconnected from the keywords on a web page that the user may have no interest in the advertisement. Yet another problem stems from using relatively small sets of keywords from web pages recently viewed by the user. Because the keywords may be selected from a small portion of the user's browse information, using these keywords may produce advertisements that do not accurately reflect the user's interests.

Certain embodiments herein describe systems and methods for providing targeted advertisements. In certain embodiments, word data and behavioral data of a plurality of users may be mined from network resources such as web sites. This data may be analyzed to detect relationships between words, categories of words, advertisements and other assets, combinations of the same, and the like. Using these relationships, in certain embodiments more targeted advertisements may be provided to a user.

As used herein, the term “asset” is used interchangeably to refer to an asset itself (for example, an advertisement, product, content, persona or profile, and the like) and to its description or representation in a data repository. As will be apparent from the context in which it is used, the term is also sometimes used herein to refer only to the asset itself or only to its representation in the data repository.

The features of these systems and methods will now be described with reference to the drawings summarized above. Throughout the drawings, reference numbers may be re-used to indicate correspondence between referenced elements. The drawings, associated descriptions, and specific implementation are provided to illustrate embodiments of the inventions described herein and not to limit the scope thereof.

In addition, methods and processes described herein are not limited to any particular sequence, and the blocks or states relating thereto can be performed in other sequences that are appropriate. For example, described blocks or states may be performed in an order other than that specifically disclosed, or multiple blocks or states may be combined in a single block or state. Moreover, the various modules of the systems described herein can be implemented as software applications, hardware and/or software modules, or components on one or more computers, such as servers. While the various modules are illustrated separately, they may share some or all of the same underlying logic or code.

illustrates an embodiment of an asset recommendation systemfor recommending assets such as advertisements to users. The asset recommendation systemcan be owned or operated by an advertisement provider or the like. In certain embodiments, the asset recommendation systemprovides targeted advertisements to one or more third party serversover a communications medium Page 5 of 36such as the Internet or other network. In turn, the third party server or serverscan provide the advertisements or other assets to one or more user systems. In certain embodiments, the advertisements may be advantageously targeted to users of the user systems, as will be described below.

The third party servermay be operated by a host of a network resource such as a web site or the like. The third party servermay, for example, serve content on one or more web pages or other network resources to the user systems. In addition to providing content to the user systems, the third party servercan retrieve or otherwise receive assets from the asset recommendation system. These assets may include, for example, one or more advertisements. Other examples of assets are described below.

The asset recommendation systemincludes one or more serversfor communicating with third party serversover the communications medium. The serversmay be web servers in certain implementations. In response to receiving requests for assets such as advertisements from the third party server, in one embodiment the serversrequest advertisements or other assets from a recommendation service.

The recommendation servicemay be a web service or the like that recommends advertisements or other assets that are targeted to a particular user or group of users. The recommendation serviceincludes one or more recommendation enginesfor generating these recommendations. In certain embodiments, each recommendation engineprovides recommendations in different ways, such as by analyzing the context of a target user's browse information or history, by analyzing the behavior of a target user, by targeting a target user demographically or geographically, by analyzing social and/or user momentum, combinations of the same, and the like. A selector modulecan determine which of the recommendations generated by the recommendation enginesto provide to the third party serverfor display to a user via a user system.

One or more of the recommendation enginesmay include components (see, for example,) for mining data from remote network resources such as web sites. These components may mine data by crawling or otherwise retrieving the data in certain embodiments. In certain embodiments, the data mined by the recommendation enginesmay include word data, user behavioral data, user tracking data, combinations of the same, and the like. Word data can include, for example, words and phrases found in the content of network resources such as web pages.

In certain embodiments, the word data is advantageously mined from social network data. In addition to having its broad ordinary meaning, “social-network data” may include data generated on social sites such as social networking sites, blogs, declared interest sites (for example, digg.com), news and other current event sites. In addition, social-network data may include web conversations, user profiles, online blogs, commentary, and the like. Word data may also be mined from reference data such as may be obtained from dictionary and encyclopedia sites, directories, search engines, combinations of the same, and the like. Because a large variety of words can be mined from such diverse network resources in certain embodiments, relationships can be detected between the word data that facilitate providing more accurately targeted asset recommendations, as will be described below.

User behavioral data may include data on browse behavior, asset selection behavior, and/or purchase behavior of one or more of users. Browse behavior data can include data on network resources accessed by users. Asset selection behavior data may include data regarding advertisements and other assets selected (for example, clicked) by users. Purchase behavior data may include data regarding online purchases of products and/or services by users. User tracking data may include data relating to a target user, who may be a user of a user systemfor whom recommended advertisements will be generated. The user tracking data for a particular user may include data regarding a recent browse or purchase history of the user, demographic data on the user, geographic data related to the user (for example, a locale in which the user lives), combinations of the same, and the like.

In certain embodiments, one or more of the recommendation enginesstore the mined data as raw data in a data repository, which may be a storage device, database, or the like. The raw data may be mined by the recommendation enginesin an off-line process or in a process that may operate in parallel or substantially in parallel with processes for providing recommendations to users. The recommendation enginesmay mine the raw data, for example, on a periodic basis (for example, Page 7 of 36 daily), or the recommendation enginesmay mine the raw data continuously or substantially continuously.

Certain components of one or more recommendation engines(see, for example,) may analyze the raw data to detect or create relationships among the raw data. The recommendation enginesmay analyze the raw data stored in the data repository. Alternatively, the recommendation enginesmay analyze raw data directly as it is being mined. The relationships created by the recommendation enginemay be stored as relationship data. Many types of relationships may be created by the recommendation engines. Word relationships may be created, for instance, between words in the word data. Other types of relationships may also be created, as is described below with respect to.

Because the word data may be mined from social sites, reference sites, and the like, the word relationships may reflect associations made on social sites, reference sites, and the like. For instance, an analysis of the word data may detect that the word “Ford®” is related to the word “barbeque” because these words are used in similar contexts by bloggers. Thus, a word relationship may be created between “Ford®” and “barbeque.”

The recommendation enginesmay use the relationship datato generate targeted recommendations. In certain embodiments, one or more recommendation enginesretrieve user tracking data for a target user from the data repository. The recommendation enginesmay analyze the user tracking data to find words in the tracking data. For example, one recommendation enginemay search for target words on a network resource that the user has recently accessed. The likely contextual meanings of the target words may be determined, for example, by the process ofdescribed below. Using the target words, the recommendation enginescan find related words in the relationship data.

In turn, using the related words, the recommendation enginescan find related assets in the relationship data. The recommendation enginescan find related assets in the relationship databy comparing the related words with keywords associated with one or more assets in the asset data. These keywords may be associated with the assets in a separate or off-line process, which may be Page 8 of 36 performed, for example, by the recommendation service. In alternative embodiments, the keywords can be manually associated with at least some of the assets. An administrator of the asset recommendation system, for instance, may analyze the assets and assign keywords by hand. Categories may also be assigned to the assets manually or automatically by inferring the categories from the assigned keywords (see, e.g.,).

The recommendation servicemay perform the keyword-asset associations by analyzing descriptive text of the assets to extract keywords. For instance, the recommendation servicemight extract keywords such as “truck” and “Ford®” from an advertisement asset that describes Ford®trucks. More details on this extraction process are described below with respect to. The keywords can also be weighted, for example, based on a part of speech of a word (for example, nouns may be weighted higher than verbs), based on a frequency of the word's use in the asset description, based on a relative position of the word in the descriptive text (for example, words higher in the descriptive text may be weighted more), combinations of the same, and the like.

By determining which words are related to the words found in the user's data, the recommendation enginemay therefore determine which assets may be candidate recommendations for the target user. The recommendation enginescan score or otherwise rank the asset recommendations based at least in part on the strength of word, asset, or other relationships in the relationship data.

The recommendation enginesmay recommend one or more assets stored in a data repository, which may include a database of assets. The data repositorycan include, for example, content related to advertisements, content describing products, content related to other users, such as user profiles, combinations of the same, and the like. Thus, while the remainder of this specification generally refers to recommending advertisements, other assets may also be recommended to users.

In certain embodiments, one or more of the recommendation enginesprovides their candidate recommendations to the selector. The selectorcan select a subset of assets to provide to the third party server, by, for example, selecting a most highly ranked or scored subset of the assets. The selectorcan provide the subset of assets to the third party server. The third party servermay output an asset for display for one or more of the user systems.

Although the third party serveris provided in the depicted embodiment, the third party servermay be removed in other embodiments. As such, the user systemsmay directly access the asset recommendation systemin certain embodiments.

Although the assets described herein are primarily referred to as including advertisements, assets may also include products (or services), content, personas, and the like. Product assets may include descriptions of products, which may be recommended, for example, to a user of a product network resource or web site that sells the given product. Content assets can include news stories, biogs, or the like, which may be recommended to users. Persona assets can include profiles of people that may be recommended, for example, on social networking sites, dating services, and the like. The remainder of this specification will refer primarily to advertisement assets; however, some or all of the systems, modules, and processes described herein may be used to provide product, content, persona, and other types of asset recommendations.

illustrates a more detailed embodiment of a recommendation engine. The recommendation engineincludes a data aggregation module, a relationship mining module, and a recommender. The recommendation enginemay use these components to mine data, analyze the data to detect relationships among the data, and generate asset recommendations based on the detected relationships.

In certain embodiments, the data aggregation moduleis configured to mine raw word and/or user data obtained from social and/or reference network resources. For example, in certain embodiments, the data aggregation modulecrawls a network such as the Internet to obtain this data. Alternatively, the data aggregation modulecan receive this data from external sources. The data aggregation modulecan store the raw data in a data repository, which may be a database or the like.

As described above with respect to, an asset recommendation system may include multiple recommendation engines. These recommendation enginesmay have the same general structure as the recommendation enginein certain embodiments, but the components of the recommendation enginesmay have Page 10 of 36 different functions. For example, the data aggregation moduleof one implementation of a recommendation enginemay gather word data but not user data, whereas a data aggregation moduleof another recommendation enginemight gather user data but not word data. Still other data aggregation modulesgather both word data and user data. More specific examples of recommendation enginesare described below with respect to.

A relationship mining modulemay access the word and/or user data stored in the data repositoryand detect or create relationships among the data. Many types of relationships may be created, including word, category, asset, and behavioral or user-based relationships.

In certain embodiments, the relationship mining moduledetects word relationships by analyzing the proximity of certain words to other words on a given network resource such as a web site. For example, if a blogger includes the words “backgammon” and “bird watching” in the same sentence or within the same passage of text, the relationship mining modulemay create a word relationship between these words. The relationship mining modulemay also create word relationships based on interactions between users with respect to one or more network resources. For example, if a social site has a user “Bob” and Bob likes video games, and if Bob has a friend Joe who likes skydiving, then the relationship mining modulemay create a word relationship between “video games” and “skydiving.” The relationship mining modulemay store the relationships in a network or graph data structure. In certain embodiments, a separate graph data structure may be created for each type of relationship. The graph data structure may be implemented in numerous ways, including but not limited to, arrays, adjacency lists, adjacency matrices, symbolic graphs, database, combinations of the same, and the like. The graph data structures may include weighted Bayesian networks, Markov networks, or the like. The graph data structures may have nodes representing words, categories, assets, or user data. A relationship between two items of data may be indicated by an edge connecting two nodes.

Referring to the “backgammon” example above and with reference to, the relationship mining modulemay construct an example graphhaving a backgammon nodeOa and a bird watching nodeOb. An edgeconnects the nodesOa andOb, indicating a relationship between backgammon and bird watching. As shown, the relationship mining modulein the depicted example has also determined that the words “coin collecting” are related to both backgammon and bird watching. Thus, a nodeOc for coin collecting is added to the graphOa, with edgesandconnecting the nodesOa andOc and the nodesOb andOc, respectively. While the graphillustrates a simple example of relationship data, much more complex graph structures can be constructed having millions or more nodesand edges.

The relationship mining modulein certain embodiments can weight or rank the relationships between words. In embodiments where the relationship dataincludes graph structures or the like, these weightings can be represented as weightings of the edgesconnecting the nodes. In certain embodiments, the weightings may be based on criteria such as frequency or popularity of certain relationships, recency of relationships, and/or authority of relationships. Frequency of relationships can refer to a number of times and/or a number of distinct sites or passages in which a relationship is detected between particular words. If many users of network resources such as social sites are describing backgammon in conjunction with bird watching, for instance, the word relationship between backgammon and bird watching may receive a higher weighting.

Recency of relationships can refer to how recently the words were published or otherwise provided on network resources. More recent relationships may be given greater weight than older relationships. If users were discussing coin collecting and backgammon within the past few weeks, for example, the weight of this relationship might be higher than if users were discussing the same words several months or years ago. Authority of relationships can refer to the authority of users or network resources from which the relationships are derived. Higher authority can result in a higher weighting. To illustrate, relationships derived from words on an anonymous comment on a blog entry might be considered to have little authority, whereas relationships derived from words in a reference work such as a dictionary might have greater authority and hence more weight.

Some or all of the frequency, recency, and authority criteria may be used to weight the relationships. In certain embodiments, the weightings of relationships dynamically change over time based on changing frequency, recency, and/or authority criteria. Thus, final recommendations generated based on the same relationships might be different at different points in time. In an embodiment, using some or all of these changing criteria may effectively weight relationships according to social trends or social momentum. Relationships that are more frequent (for example, popular), recent, and/or which have more authority might reflect a degree of social momentum that the relationship has. In certain embodiments, recommendations can be generated based on relationships having greater social momentum than other relationships, resulting in timely, targeted recommendations.

As one example of weighting word relationships, authority of social sites may be first determined manually by an administrator. The administrator might determine, for instance, that blogging sites are generally more credible or have more authority than other social sites such as My Space®, Facebook®, or the like. A sample set of social network data may be taken based on the activity of users of those sites. The sample set of social-network data might include, for example, data created by between 10,000 and 50,000 users. These users may be selected based at least in part on the level of activity on their respective social sites. This level of activity corresponds to the recency factor in certain embodiments. With the sample set of social-network data obtained, word data can be extracted from the social-network data and word relationships may be created as described above. The frequency of the occurrence of words within the social-network data may be used to further weight the word relationships.

As social momentum can change over time, relationships may be strong at one point in time and then become weak, or vice versa. For example, if many users of social sites, news sites, or the like are talking about backgammon and bird watching, social momentum might be high. Years, months, or even days later, users may lose interest in these topics and the social momentum might decrease significantly. Recommendations can be automatically adjusted accordingly in certain instances to provide fewer recommendations related to relationships with decreased social momentum. For example, the weightings of the relationships may be aged or decreased, so that more Page 13 of 36 recent relationships may have higher weightings, which may result in more recent relationships being used to generate recommendations.

In addition to detecting word relationships, in certain embodiments, the relationship mining moduleofcan also detect and create category relationships. In certain embodiments, the relationship mining moduleinfers category relationships from word relationships. To do this, the relationship mining modulemay access category data from the data repository. This category data may be obtained by the data aggregation moduleby, for example, using linguistic categorizations, which may be obtained from online dictionaries, such as wordnet.princeton.edu, or other like online reference resources. The data aggregation modulemay also obtain the category data from societal categorizations, which may be obtained from directory sites that arrange topics in hierarchical category form, such as dmoz.org. Creating relationships between categories can allow generalized concepts to be related. When generating recommendations, the recommendation enginecan use the category relationships to find related assets if no assets exist that are related to a word relationship.

Referring to, categories for the words inare shown with inferred category relationships. Thus, a “board game” node, a possible category for backgammon, is shown related to a “naturalist” node, a possible category for bird watching. Likewise, a “collecting” node, a possible category for “coin collecting,” is shown related to the nodes,. Edges,, andconnect the nodes,, and

Referring again to, the relationship mining modulemay also infer asset relationships from category relationships. To do this, the relationship mining modulemay access asset data from a data repository. This asset data may include keywords for each asset, as described above with respect to. In addition, the asset data can include one or more categories for each keyword, which may be obtained, for example, by analyzing reference data in a similar manner to the category inferences described above. To infer asset relationships from category relationships, the relationship mining modulemay determine which categories in the category relationships overlap with categories of assets in the asset data.

For example, with reference to, an inference from a word relationshipto a category relationshipis illustrated by arrow. In addition, the inference of an asset relationshipfrom the category relationshipis illustrated by arrow. In the depicted example, the assets include an advertisement for a “chess game” at node, which can be inferred as being related to the category “board game.” The assets further include an advertisement for a “hiking book” at node, which can be inferred from the category “naturalist.” An edgeconnects the nodes,, reflecting the inferred relationship between the chess game advertisement and the hiking book advertisement.

Turning again to, the relationship mining modulemay also generate asset relationships directly. Asset relationships may be created by analyzing user browse data and/or reference data. As an example of analyzing user browse data, the relationship mining modulemay relate two product assets because a user viewed one product and then soon after viewed the second product. As an example of analyzing reference data, the relationship mining modulemay relate an advertisement asset to a content asset because both the advertisement and the content contain the same or similar dictionary words.

From these directly created asset relationships, the relationship mining modulemay also infer word relationships from the assets and in turn infer category relationships from the word relationships. The relationship mining modulemay infer word relationships by retrieving words in the related assets. Once the word relationships are created, the relationship mining modulemay create the category relationships. For example,illustrates that an asset relationshipmay be created between an advertisement for a backgammon game at nodeand a bird watching book at. Arrowindicates that an inference may be made that the word “backgammon” at nodeOa is related to the words “bird watching” at nodeOb. Arrowfurther indicates that an inference may be made that the category “board game” at nodemay be related to the category of “bird watching” at nodeOb.

Turning again to, a recommenderof the recommendation enginemay use the relationship datato generate targeted recommendations. In certain embodiments, the recommendercan retrieve user tracking data for a target user from the data repository. The recommendermay analyze the user tracking data to find words in the tracking data and find related words in the relationship data. The recommendermay then use the related words to find related assets in the relationship data. The recommendercan recommend one or more of these assets for a target user. The recommendermay also use other recommendation generation techniques, some examples of which are described below with respect to.

FIG. SA illustrates an embodiment of a relationship creation processfor creating word relationships based on word data obtained from various network resources. The processmay be implemented by the asset recommendation systemor by the recommendation enginein certain embodiments. The processfacilitates detecting relationships between words that can be used to generate recommendations for users.